Séminaire statistique

Description :

Conférencier :   Pankaj Bhagwat (University of Alberta)

Titre : From Models to Reliability: Improving Conformal Prediction through Bayesian Model Averaging

Résumé :

Conformal prediction is a versatile technique that ensures valid predictive inference across various machine learning models by generating prediction sets with finite-sample marginal coverage guarantees. Unlike traditional methods that rely on strict distributional assumptions, conformal prediction is distribution-free, adapting predictions from any model into reliable intervals. Typically, it involves selecting a machine learning model and generating prediction sets with a specified coverage level. However, with a wide range of potential models for any problem, identifying the most suitable one for optimal prediction sets is challenging. In this seminar, I will explore this challenge and present a novel approach that integrates conformity scores from multiple models through Bayesian model averaging. This hybrid method enhances the strengths of conformal prediction with the added robustness of model averaging.  Theoretically, we demonstrate that the resulting prediction interval converges to the optimal accuracy level if the true model is among the candidates. Moreover, our method incorporates both data and model uncertainty into prediction interval construction, enhancing the reliability of predictions.